acceptodds
Under review as a conference paper at ICLR 2027

Tropical Diffusion Dynamics: Why Nearly Identical Diffusion States Reach Different Output Routes

Abstract

This paper investigates how deterministic probability-flow ordinary differential equations in diffusion models can produce different terminal outputs from nearly identical intermediate states. We address this by framing Gaussian denoising as a soft-minimum competition among clean-data routes, whose zero-noise limit yields nearest-support cells and, for finite atomic priors, tropical power cells. We prove that isolated regular two-route boundaries form a logistic transition layer of width with peak denoiser sensitivity , where is the noise standard deviation. We show that between-route posterior covariance simultaneously determines the Fisher tensor of route probabilities, the route component of reverse-flow strain, and the rate of Gaussian route-information gain. Because the remaining flow transports and amplifies these contrasts before decoding, we introduce a matrix-free Cauchy–Green Jacobian–vector product probe that maximizes linearized terminal displacement within a low-rank subspace. These theoretical insights are operationalized into Tropical Route Control (TRC), a modular framework combining finite-horizon probes, trajectory-level conformal event control, and exact prefix allocation under a hard continuation budget. Our experiments verify the predicted local sensitivity and finite-horizon perturbation effects and demonstrate that route-aware verification provides a valid and practical way to detect declared endpoint changes under a fixed computation budget in real settings.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion models, tropical geometry, diffusion dynamics, probability flow ODEs

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Zw4DOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies why nearby intermediate states can produce different outputs under deterministic diffusion sampling. It describes Gaussian denoising as competition among clean-data routes and connects posterior covariance to route sensitivity, flow strain, and information gain. It then introduces a finite-horizon Jacobian probe and a Tropical Route Control framework combining event calibration with budgeted verification. Controlled experiments and DiT/SDXL studies support several sensitivity and event-capture observations, but the connection between the tropical route theory and the practical verification pipeline remains only partially established.

Critical strengths
  1. A useful link between local uncertainty and changes in the final output. The paper explains why a state that is sensitive to small changes now may not produce the largest change in the final image: later sampling steps can amplify or reduce those changes. The analysis makes this distinction clear and states the assumptions and limits of its results.
  2. Extensive controlled experiments and transparent reporting. The experiments include matched perturbation controls, independent trajectory cohorts, multiple solvers, and an exploratory SDXL transfer study. The finite-horizon direction outperforms a matched random direction on all 64 trajectories in the main DiT test, while risk-based allocation improves event capture under fixed continuation budgets. The appendix also reports negative results and failed validation checks rather than presenting uniformly favorable outcomes.
Critical weaknesses

The practical event is not sufficiently connected to the theoretical route geometry. The deployment target is thresholded CLIP/pixel dispersion, whereas the route-boundary theory concerns specified mixture components or fixed terminal cells. Table 18 reports only approximately 61% agreement between the dispersion and classifier-cell events, and Table 17 does not establish the predicted small-radius scaling regime. These observations do not contradict the conditional theorems, but they leave unclear when tropical route competition, rather than generic within-route variation, explains the measured events in large models.

The main allocation results do not establish the added value of the integrated TRC framework. Appendix D.1 states that deployment uses Gaussian probe directions and a gradient-boosted risk predictor with neither terminal JVPs nor directly measured between-route covariance. Thus, the strongest capture results do not test the proposed finite-horizon direction module. Moreover, Table 18 does not establish superiority of the certified Band-TRC allocator over uniform allocation. The individual components are reasonably motivated, but their combined benefit beyond a generic learned event-priority policy remains un demonstrated.

The practical value of finite-horizon probes under matched total computation is unclear. The main direction experiment evaluates terminal latent gain rather than the deployed decoded-feature event. Table 14 reports approximately 10.38 seconds for exact rank-8 direction construction, versus 0.92 seconds for the instantaneous alternative and 0.41 seconds for one signed continuation pair. Finite differences substantially reduce construction cost, but the experiments do not establish which probing strategy is preferable under an equal total runtime budget.

Questions for authors

Please refer to weaknesses.

6 comments

We sincerely thank the reviewer for the careful and insightful assessment. We particularly appreciate the recognition of the distinction between instantaneous sensitivity and finite-horizon terminal sensitivity, as well as the rigor of our controlled sanity check.

We would like to clarify the central contribution and intended scope of our work. Tropical Diffusion Dynamics is primarily a theory-driven contribution. To our knowledge, it provides the first unified theoretical framework connecting competition among plausible clean-data routes, the resulting tropical boundary geometry, the role of between-route posterior covariance in reverse-flow strain and information acquisition, and the propagation of local sensitivity through deterministic diffusion dynamics to terminal outputs.

Building on these theoretical findings, we introduce Tropical Route Control (TRC) as an initial operational framework for finite-horizon probing, statistically controlled endpoint verification, and budgeted computation allocation. Our intention is not to present TRC as a replacement for existing state-of-the-art sampling algorithms, but to demonstrate how the proposed theory can provide a principled foundation for analyzing and verifying terminal sensitivity. We agree that the connection between the theoretical framework and its practical operationalization deserves further clarification. We have conducted additional targeted experiments and elaborate on these connections below.

 

1. Concern: When does the measured event reflect tropical route competition rather than generic within-route variation?

A1. We appreciate the reviewer for identifying this important distinction. We agree that three mathematically different objects should be distinguished:

  1. Clean-data route posterior: the posterior responsibilities of specified clean-data routes under Gaussian corruption.
  2. Transported terminal decision partition: a specified terminal decision boundary pulled back to an intermediate state through the remaining deterministic sampling flow.
  3. Finite-cloud endpoint-dispersion event: an operational event defined by whether outputs from a finite perturbation cloud exceed a prescribed feature-space distance threshold.

These objects are related by our theoretical framework, but they are not interchangeable.

Regarding Table 18, we agree that feature-space dispersion and classifier-defined decision changes do not generally coincide. In two independent cohorts, the exact partitions agree in 58.0% and 58.5% of cases, while their split/no-split decisions agree in 61.6% and 60.9%, respectively.

We believe this discrepancy is informative and consistent with an important distinction established by our theory.

Theorem 2 gives the exact posterior covariance decomposition:

Σq=Σqwithin+Σqbetween.\Sigma_q=\Sigma_q^{\mathrm{within}}+\Sigma_q^{\mathrm{between}}.

The within-route component describes posterior variation inside individual routes, whereas the between-route component captures competition between route-conditioned posterior means. In particular, the between-route component determines the spatial Fisher information of the specified route responsibilities through

Σqbetween=q2IR.\Sigma_q^{\mathrm{between}}=q^2 I_R.

Consequently, an output-dispersion event may arise from either within-route variation or between-route competition. A feature-distance threshold alone cannot distinguish these mechanisms.

Theorem 26 further characterizes this distinction geometrically through its bulk–wall decomposition: within-route responses contribute to the absolutely continuous bulk, whereas between-route susceptibility concentrates on regular interfaces of the route skeleton, under the stated projection regularity and exceptional-set assumptions.

Thus, the incomplete agreement in Table 18 does not contradict our theoretical results. Rather, it illustrates why output dispersion should not automatically be interpreted as a semantic route transition, and why distinguishing within-route variation from between-route susceptibility is important when connecting the theory to real models.

To further examine whether these sensitivities correspond to meaningful generation changes, we conducted an attribute-level evaluation on native SDXL.

Using 553 GenEval prompts, we evaluated the original output and eight perturbed continuations for each prompt with the official attribute evaluator.

Attribute-level observationPrompts
At least one output passes and another fails91/553
Opposite perturbations along one direction give different decisions85/553
Original feature threshold also crossed among attribute-changing prompts5/91

These results demonstrate that small perturbations of intermediate states can change requested generation attributes in a real pretrained diffusion model. Notably, only five of the 91 attribute-changing prompts also crossed the original CLIP/pixel feature threshold. This further illustrates that feature-space dispersion and task-level semantic changes are different observables.

We also investigated 20 cases with previously observed attribute flips by scanning nine locations along the identified direction and nine locations along a matched orthogonal direction. Across 360 local signed-pair tests, 23 exhibited different attribute decisions.

We then evaluated how well different directional responses rank these local attribute changes:

Directional responseAUC
Denoiser response0.683
Terminal CLIP response0.501
Terminal latent response0.497

The denoiser response provides a more informative ranking signal than the two terminal-distance alternatives in this exploratory scan. This offers additional evidence that intermediate-state sensitivity contains information about subsequent task-level output changes.

These results establish an important connection between our theoretical framework and the behavior of real diffusion models. The attribute-level evaluation provides direct evidence of terminal decision boundaries in pretrained SDXL, while the directional scans further reveal how these boundaries are localized in the intermediate state space. This is precisely the type of geometric structure captured by our transported-boundary analysis. Furthermore, the posterior route formulation provides a probabilistic interpretation of the competing clean-data explanations underlying such transitions. Together, these experiments demonstrate that the terminal sensitivity and boundary phenomena studied in our theory have observable and practically meaningful counterparts in large-scale diffusion generation.

Regarding Table 17 and the small-radius scaling law, we agree that the observed finite-radius experiments do not establish the predicted asymptotic exponent of one. The theoretical result is conditional on a regular transported boundary and the limit as the perturbation radius approaches zero. The tested radii in pretrained models are not known to lie within this asymptotic regime.

Importantly, our theoretical analysis also develops non-asymptotic results that directly address finite perturbation radii.

Theorem 5 provides a finite-radius lower bound on the symmetric terminal gain along the leading direction:

gρ(v⋆)≥ς1−L3ρ26.g_\rho(v^\star)\geq\varsigma_1-\frac{L_3\rho^2}{6}.

Theorem 19 further characterizes the diameter of the entire symmetric endpoint cloud, including endpoint pairs beyond a single signed perturbation:

Dρ=2ρA+O(ρ3).D_\rho=2\rho A+O(\rho^3).

Both results hold under their respective smoothness, rank, and nondegeneracy assumptions.

Finally, Theorem 6 establishes trajectory-level statistical control for a pre-specified finite-cloud event under exchangeability. Importantly, this guarantee does not require the empirical event prevalence to follow the asymptotic tube-scaling law.

Therefore, the finite-radius deviations do not invalidate the conditional theoretical results or the separate statistical guarantee used by TRC.

 

2. Concern: What does the integrated TRC framework add beyond a generic learned event-priority policy?

A2. We appreciate this important question. We would first like to clarify that the three TRC components address different aspects of the verification problem:

  • Finite-horizon probing determines which perturbation directions are informative about subsequent terminal sensitivity.
  • Trajectory-level event calibration provides statistical control for a pre-specified endpoint event.
  • Budget allocation determines where limited verification computation should be spent.

These components are complementary, but their empirical contributions should be evaluated separately rather than assuming that their combination automatically yields a superior sampling algorithm.

More broadly, the objective of TRC is orthogonal to the primary objective of conventional diffusion samplers. Existing numerical solvers optimize trajectory integration, while covariance-aware sampling and endpoint-prediction methods exploit posterior or endpoint information to improve sampling efficiency and generation quality. TRC instead asks whether controlled perturbations of intermediate states lead to specified terminal changes, and how limited computation can be allocated to verify such events.

We therefore position TRC as a theory-grounded verification framework and an algorithmic sanity check of the proposed sensitivity mechanism, rather than a replacement for existing state-of-the-art samplers.

To investigate the individual contributions of direction selection and budget allocation, we conducted a factorial comparison crossing three probe directions with two allocation policies on the same 256 DiT-XL/2 trajectories. All probe-construction costs were included in the computational budget.

BudgetAllocationRandomTerminal latentDecoded feature
50%Uniform929292
50%Risk priority102108107
75%Uniform112121120
75%Risk priority112121120

These comparisons help separate the two design choices.

At the 50% budget, changing the allocation policy from uniform to risk priority increases the number of recovered feature events from 92 to 102 for random probes. With risk priority fixed, using terminal-latent or decoded-feature directions further increases the count to 108 and 107, respectively.

At the 75% budget, changing the probe direction under uniform allocation adds eight events for decoded-feature directions and nine for terminal-latent directions relative to random probes. The two allocation policies produce the same counts in this cohort.

We regard these results as evidence that direction selection and computation allocation provide distinct and potentially complementary sources of information. We do not claim a universal additive gain or statistically established synergy between the two modules. We also clarify that the original event-priority allocation results should not be attributed directly to the finite-horizon Jacobian probe: the factorial experiment is intended to examine the contribution of direction selection separately.

The certified Band-TRC allocator serves a different purpose from empirical risk-priority scheduling.

CohortFull-cloud positivesBand-TRCUniform
Prospective117102100
Independent replication135120119

The observed capture differences are modest: two additional positives in the prospective cohort and one in the independent replication. We therefore do not claim empirical superiority of Band-TRC over uniform allocation. Its primary contribution is the statistical guarantee under the stated assumptions, rather than maximizing the raw number of captured events.

We believe this framework has several potential applications beyond event capture itself.

First, reliability-oriented generation and verification: in scientific visualization, reconstruction, or autonomous-driving simulation, users may require particular output attributes to remain stable under small perturbations of intermediate states. TRC could identify trajectories requiring additional verification, conservative continuation, resampling, or human review.

Second, solver robustness benchmarking: conventional generation metrics primarily assess average output quality, whereas TRC provides a complementary way to evaluate the sensitivity of trajectories to controlled perturbations under a specified terminal criterion.

Third, failure analysis and mechanistic understanding: the framework can help identify when a specified terminal decision becomes sensitive along a sampling trajectory and whether subsequent dynamics amplify or suppress this sensitivity.

Under the exchangeability and calibration assumptions of Theorem 6, TRC provides event-specific trajectory-level false-acceptance control. This is a theoretical advantage for selective verification, although it should not be interpreted as a universal guarantee of semantic correctness or generation safety.

The current experiments provide an initial demonstration of this direction. Developing stronger algorithms that jointly account for route-sensitive probing, task-level correctness, and the expected benefit of additional computation is a promising direction for future research.

Reviewer_R673OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses a core phenomenon in deterministic diffusion models: nearly identical intermediate states can diverge into drastically different outputs after deterministic denoising. The authors frame Gaussian denoising as soft-minimum competition among multiple clean data routes, and employ a tropical geometry framework to characterize the structure and evolution of route boundaries. They prove that regular two-route boundaries exhibit a logistic profile, with width proportional to noise variance and peak sensitivity inversely proportional to noise variance. The work further shows that the between-route posterior covariance tensor serves three simultaneous roles: quantifying the spatial sensitivity of route probabilities, constituting the route component of reverse probability flow strain, and determining the rate of route information gain in Gaussian channels.

Building on these theoretical findings, the authors propose the Tropical Route Control (TRC) framework, which enables endpoint change detection and verification under a fixed compute budget via three modules: finite-horizon Jacobian probes, trajectory-level conformal event control, and dynamic programming for budget allocation. For validation, the authors conduct controlled mechanism verification on small U-Net and lightweight DiT models, as well as phenomenological validation on large-scale DiT-XL/2 and native SDXL models.

This work proposes a novel theoretical perspective and provides a complete tropical geometry-based explanation for the route divergence phenomenon in diffusion models. The theoretical derivations are rigorous, and the controlled experiments are solid. However, the work currently remains primarily theoretical, with insufficient large-scale validation and unclear applied value. The overall presentation also poses notable reading difficulties. Taken together, there remains a gap between this work and the acceptance standard for ICLR.

Critical strengths
  1. This work breaks away from the engineering-optimization dominated mainstream paradigm in diffusion model research. It explains the route divergence phenomenon in diffusion processes from a novel tropical geometry perspective, and establishes a complete theoretical chain from local boundary geometry to terminal output evolution. This makes an important original contribution to understanding the fundamental problems of diffusion dynamics, and represents the core value of the paper.

  2. The controlled experiments on small U-Net and lightweight DiT models follow a rigorous design. By strictly matching perturbation norms, subspaces, and sampling steps, the authors verify theoretical predictions through causal interventions, ROC analysis, and ablation studies. The conclusions at the mechanism level are solid and reliable, providing a reference validation paradigm for similar theoretical studies.

Critical weaknesses
  1. The paper presents TRC as a practical deployable framework, but existing experiments fail to demonstrate its effectiveness and advantages in real scenarios. On one hand, large-scale experiments only compare against weak baselines, without benchmarking against existing state-of-the-art methods for related tasks. On the other hand, the small sample size of the SDXL experiment limits the statistical confidence of cross-architecture generalizability.
  2. Core tropical geometry concepts are built on ideal Gaussian mixture assumptions. The paper does not verify the existence of corresponding soft tropical cells and boundary structures in real trained diffusion models, nor does it provide direct visualization or quantitative characterization of the tropical skeleton in real models. The practical relevance of the theory thus lacks direct evidence, and risks being dismissed as a mathematical conclusion only valid under idealized settings.
  3. The entire work remains at the level of phenomenon explanation and theoretical verification. There is no clarification of real-world deployment scenarios for the TRC framework, such as whether it can improve sampling quality, assist generation content editing, detect generation faults, or enhance model robustness. The impact and practical value of the work remain unclear.
Questions for authors

Apart from what is presented in the "Critical Weaknesses" section, I also have some secondary concerns:

  1. The main text is overly dense and difficult to follow, with insufficient supporting illustrations. The experimental section in the main text is too thin to fully convey the findings.
  2. The custom "route divergence event" metric lacks consensus in the field. The rationale for threshold setting is not sufficiently justified, reducing the interpretability and comparability of experimental results.
  3. The related work section does not discuss existing work on diffusion geometry, covariance-aware sampling, and endpoint prediction in sufficient depth, and fails to clearly distinguish the core differences and marginal contributions of this work from prior studies.
  4. The AI Use Statement, Reproducibility Statement, and Ethics Statement are placed in an inappropriate position. Per standard conference formatting requirements, they should appear after the main text and before the reference section.
  5. The correspondence between core theoretical concepts (soft tropical cells, route boundary layers, tropical skeleton) and the generation process of real diffusion models is not clearly contextualized in the main text, which may lead application-oriented readers to perceive the mathematical derivations as disconnected from practical generation.

Therefore, I suggest that the author make the following revisions:

  1. Include existing state-of-the-art methods as baselines in DiT-XL/2 experiments, such as covariance-aware sampling and endpoint prediction-based verification methods, to explicitly demonstrate the relative advantages of the TRC framework under matched compute budgets.
  2. Supplement direct observational experiments of tropical boundary structures in real diffusion models. For example, use hidden state clustering and boundary detection to verify the existence of soft tropical cells, establishing a direct link between theory and real models.
  3. Add validation of at least one practical deployment scenario, such as using TRC for guided generation editing or sampling fault detection, to demonstrate the practical value of the work.
  4. Move key proof sketches and experimental metric descriptions into the main text, add necessary illustrations and explanations to reduce reading burden. Clearly distinguish the scope of theoretical conclusions and experimental conclusions to avoid over-interpretation.
8 comments

We sincerely thank the reviewer for the thoughtful and constructive assessment. We are particularly encouraged by the recognition that our work establishes a novel tropical-geometric perspective on diffusion dynamics, together with a rigorous theoretical connection from local route competition to terminal output sensitivity. To our knowledge, this is the first theoretical framework to provide a unified and comprehensive explanation of the route-divergence mechanism in diffusion models. We also appreciate the positive assessment of our controlled experiments.

We would like to clarify one important aspect of our intended contribution: this is primarily a theory-driven paper, rather than an attempt to introduce another state-of-the-art sampling algorithm. Our main objective is to establish a unified mathematical framework that explains how competition between plausible clean-data routes induces local sensitivity, how this sensitivity evolves through the remaining deterministic sampling dynamics, and how it can ultimately affect terminal generation outcomes.

TRC is an initial operational framework motivated by this theory. Its purpose is to investigate how theoretically characterized sensitivity can support risk-aware probing, endpoint verification, and computation allocation. We agree that the current implementation is not yet a competitive replacement for existing samplers, and we will clarify this distinction throughout the paper.

Thank you for asking the valuable questions. We hope the replies below, with more detailed explainations and new experiments, make the answers clear.

1. Concern: Practical effectiveness, stronger baselines, and large-scale validation

A1. We appreciate the suggestion to evaluate TRC against stronger baselines and under more realistic computational budgets. We have investigated this from two complementary perspectives: generation-level computation allocation and verification-level efficiency.

 

Generation-level evaluation.

On 517 held-out GenEval prompts using native-resolution SDXL, we examined whether a simple sensitivity-related signal could help decide which trajectories should receive additional sampling computation.

After seven evaluations, trajectories with larger disagreement between two clean-image predictions received twelve additional evaluations, while the remaining trajectories returned their current estimates. We compared this adaptive allocation against random assignment with the same two computational branches and the same average evaluation budget.

Sampling ruleMean evaluationsGenEval ↑CLIP cosine ↑
Adaptive 7 or 1912.9944.16%0.2837
Random 7 or 1912.9942.10%0.2865
Fixed endpoint decoding (TJS)1347.34%0.2911
DDIM1349.61%0.2747
DPM-Solver++1350.86%0.2758
UniPC1351.30%0.2744
UniPC, lower budget1248.09%0.2727

The drift-based allocation obtained a higher GenEval score than random allocation under the matched budget (44.16% versus 42.10%), although CLIP cosine was slightly lower. We acknowledge that this simple strategy does not outperform established solvers such as UniPC or DPM-Solver++.

We do not intend to present this preliminary heuristic as a competitive new sampler. Rather, it provides an initial test of whether the sensitivity information studied in our theory can inform trajectory-dependent computation allocation.

An important observation is that predicting whether a trajectory may produce a different output is not equivalent to predicting whether additional computation will improve the requested output attributes. Developing algorithms that explicitly account for both quantities is an interesting direction for future work.

We also explored alternative uses of additional computation on 800 GenEval2 prompts, comparing extra solver steps with spatially targeted conditioning.

SDXL configurationSoft-TIFA-GM ↑Mean seconds/image ↓
APG, single output (base budget)0.09821.240
APG, extra solver steps0.09071.442
Uniform phrase conditioning0.10441.395
Regional phrase conditioning0.10731.395

Regional conditioning showed a modest improvement over uniform phrase conditioning at essentially the same evaluation count.

 

Verification-level evaluation.

We also expanded the DiT-XL/2 experiments to compare probe selection and verification under matched computational budgets, as detailed in our response to Question 8.

For the original SDXL budgeted verification experiment, the observed numbers were 10/13 positives recovered by risk-priority allocation versus 2/13 by uniform allocation. We acknowledge that this positive sample is small and treat it as preliminary evidence rather than a definitive estimate of cross-architecture performance.

The additional SDXL studies on 517 generation prompts and 553 attribute-evaluation prompts provide complementary evidence about generation sensitivity, but they should not be conflated with the statistical strength of the original budgeted verification experiment.

 

Overall, we view these results as an initial demonstration of the potential of theory-guided verification and computation allocation. We believe that further leveraging this theory to design more advanced algorithms could lead to improved performance in more downstream applications. Although such algorithmic development is beyond the scope of the present paper, we do consider it a promising and interesting direction, and an important part of our future work.

2. Concern: Can we see a boundary in a trained model, and how much of the theory depends on a Gaussian mixture?

A2. On trained SDXL we can locate the boundary of a decision about a real image. Take a prompt asking for four handbags. Holding the prompt and remaining sampling steps fixed, move the intermediate state along a line. On one side the official evaluator accepts the count. On the other it detects just three bags. Here is how often a decision changed in the full panel of 553 prompts and 4,977 outputs.

Question about the nine outputs for one promptPrompts
At least one output passes and another fails the attribute test91/553
Opposite perturbations along one direction give different answers85/553
Feature threshold crossed among prompts with an attribute change5/91

To locate individual crossings, we selected 20 cases with an observed attribute change and scanned nine points on the earlier flip direction and nine on an orthogonal line. At each point a smaller symmetric pair tested the local decision.

Scan direction (20 selected cases)Cases crossing an attribute boundaryLocal signed pairs with different decisions
Previously observed flip direction20/2015/180
Matched orthogonal direction13/208/180

The twenty scans prepared for the anonymous supplement show these crossings directly. They map the pullback of a specified output decision.

 

Regarding the Gaussian-mixture assumption, we would like to clarify its role and further elaborate on the generality of our theoretical framework.

First, Gaussianity is assumed for the forward corruption process, not for the underlying clean-data distribution. Specifically, we consider the standard Gaussian observation model Yq=X+qZY_q = X + \sqrt{q}Z, where X∼Π0X \sim \Pi_0 can follow an arbitrary probability distribution with finite second moments, and Z∼N(0,Id)Z \sim \mathcal{N}(0,I_d) is independent Gaussian noise. This is not an assumption that the clean-data distribution itself follows a Gaussian mixture. To characterize route competition, we introduce a finite latent route variable RR, such that Π0=∑a=1KπaΠa\Pi_0 = \sum_{a=1}^K \pi_a \Pi_a, where each route-conditional distribution Πa\Pi_a can be arbitrary. The Gaussian kernel is chosen because it is precisely the corruption mechanism underlying standard diffusion models, and it enables exact posterior differentiation identities.

Reviewer_2KgyOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper studies why two very close intermediate states in a diffusion model can lead to visibly different final outputs under deterministic sampling.
    • The authors propose Tropical Diffusion Dynamics, a theoretical framework that links the competition among latent clean-data routes in Gaussian denoising to tropical geometry, and analyzes the relationship between posterior covariance, local sensitivity, and sampling dynamics.
    • Building on this, they introduce TRC, which combines finite-horizon perturbation directions, trajectory-level risk calibration, and budget allocation to prioritize, under a limited compute budget, trajectories likely to exhibit large output changes.
Critical strengths
  • The theoretical framework is fairly complete. It connects tropical geometry, posterior covariance, and diffusion dynamics to explain how local perturbations can affect the final output, and the finite-horizon analysis further accounts for the cumulative effect of a perturbation propagated through the remaining sampling steps.
  • The experiments span several levels. Beyond synthetic validation, the authors compare different perturbation directions on controlled neural networks and on a pretrained DiT-XL/2.
Critical weaknesses
  • The theoretical mechanism is not yet adequately connected to semantic changes in real large models. The theory concerns route competition and decision boundaries, but the TRC experiments mainly detect output-cloud fragmentation events defined through CLIP and pixel distances. The appendix (Table 18) shows that these events exactly match ResNet-50 top-1 cell splits only about 58% of the time, which does not adequately support the claim that they reflect semantic route changes. I appreciate that the authors report this audit candidly, but the title and abstract suggest a broader explanatory scope than the evidence supports.
  • The practical use of the framework is limited. TRC is evaluated on detecting a declared output-fragmentation event under a verification budget, and the paper explicitly does not address sample quality. Showing that the explanation can translate into a concrete benefit for generation would substantially strengthen the paper.
Questions for authors
  • Is there additional evidence that the proposed mechanism explains semantic changes in real large models, for example under alternative event definitions or with attribute-level or human-judged evaluation?
  • Could the framework be used to improve generation quality? A simple idea is adaptive sampling: instead of allocating computation uniformly across trajectories and timesteps, prioritize trajectories with high predicted risk and give them more sampling steps. I note that Appendix C.4 explores event-refined Heun in low-dimensional settings.
    • Have the authors tried this on DiT-XL/2 or SDXL and measured generation quality (e.g., FID, CLIP score)?
3 comments

We appreciate your observation that our innovative perspective connects a local perturbation to what happens in the final image. Actually, this connection is the central theoretical contribution of our paper, and this is the first theory to explain this mechanism. While empirical work would be interesting, demonstrating these performances experimentally would require a substantial engineering effort and domain‑specific datasets that lie outside the paper's scope since our primary contribution is intentionally theoretical. We focused on rigorous analysis and proofs in this paper.

Thank you for asking the valuable questions. We hope the replies below, with new experiments, make the answers clear.

 

1. Concern: How closely does the event in Table 18 track a semantic change?

A1. In Table 18, we grouped the same images in two ways: by their feature distances and by their ResNet-50 labels. Asking whether the two rules produce exactly the same groups is a stricter test than asking whether they agree that a split occurred.

Table 18 comparisonCohort ACohort B
Intermediate states tested576576
Exactly the same partition58.0%58.5%
Same split / no-split decision61.6%60.9%

The distinction really matters to the theory itself. Theorem 2 separates variation within a clean route from variation between route means. Both affect the denoised estimate. Only the between-route term governs how sensitive the route probabilities are to the observed state. A change in lighting, for example, could move two images apart in feature space while leaving the object label unchanged. Equally, a small change near a decision boundary could switch that label.

This suggests a direct test of your semantic question: name the attribute first, then check whether opposite perturbations give different answers. On SDXL that happened for 85 of 553 GenEval prompts (A3). We will use “feature dispersion” for the distance event in the paper and identify the decision explicitly when reporting an attribute change.

 

2. Concern: What practical benefit follows from the theory?

A2. A sampler developer can use the theory to ask a question that average image quality leaves open: at what stage can a small perturbation change a particular outcome? To our knowledge, this is the first account to connect competition among clean routes at a tropical boundary to reverse-flow strain and to follow a perturbation through the remaining sampler. The U-Net and DiT interventions test the predicted directions. Tropical Route Control (TRC) is an algorithmic sanity check for the same idea, with a verification objective orthogonal to FID or average prompt alignment.

  • Consider a driving simulation that calls for a pedestrian at a crossing. Two images can both look realistic even if the pedestrian disappears in one of them. A generator could scan for a state sensitive to that requirement, then resample, switch solvers, or impose the object constraint there. The same question arises for a structure in a medical reconstruction or a property in a materials visualization. Theorem 6 bounds the risk of declaring stability when the specified event occurs, given separate calibration data and exchangeable trajectories.

  • A use we can demonstrate right now is comparing solvers. We perturbed the same 64 DiT trajectories at the first and last tested states and counted how often the endpoints crossed the same feature threshold.

Solver on the same 64 DiT trajectoriesEarly eventsLate events
DPM-Solver++51/641/64
DDIM46/640/64
UniPC16/640/64

The three solvers look similar at the last state and quite different at the first. That will be useful information for anyone deciding where to examine a sampler.

  • A third use is diagnosing abrupt changes: tracing the width and location of a sensitive region can identify a stage at which to try a constraint or regularizer. We plan to extend the theory to downstream tasks by using those measurements to choose and test interventions.

We also tested whether the diagnostic helps spend a fixed verification budget. On 256 DiT-XL/2 trajectories, both policies paid for the common sampling prefix, their probes, and their continuations.

Direction and allocationEvents foundModel evaluations
Random, uniform9219,820
Decoded output, risk priority10719,820

The joint policy found 15 additional events with the same number of model evaluations.

  • A4 will takes up your further question of whether extra computation improves the generated image.

 

3. Question: Can an attribute-level evaluation show such changes in a real model?

A3. Yes. We held the SDXL prompt and sampler fixed and changed only the intermediate state. For each of the 553 GenEval prompts, its official evaluator checked the original image and eight perturbed continuations against the requested objects, count, color, or spatial relation.

Question about the nine outputs for one promptPrompts
At least one output passes and another fails the attribute test91/553
Opposite perturbations along one direction give different answers85/553
Feature threshold crossed among prompts with an attribute change5/91

For 91 prompts, the perturbations changed whether the request was satisfied. Only five of those prompts also crossed the original feature threshold. This is exactly where testing the requested attribute pays off: most of these changes would be missed by a distance-only check.

We then followed the change along a line in state space. We selected 20 cases with a previously observed attribute flip, scanned nine locations along that direction and nine along a matched orthogonal direction, and tested a smaller pair of opposite perturbations at each location.

Scan direction (20 selected cases)Cases crossing an attribute boundaryLocal signed pairs with different decisions
Previously observed flip direction20/2015/180
Matched orthogonal direction13/208/180

One prompt asks for four handbags. On one side of the crossing, the image satisfies the count requirement. On the other, the evaluator finds three bags. This is the boundary in Theorem 4 made observable: after the same remaining sampling steps, neighboring states receive different terminal decisions.

Taken together, these experiments provide complementary evidence for the terminal-boundary perspective developed in Theorem 4. The large-scale evaluation establishes that attribute-level changes are not isolated cases but occur across a substantial number of real SDXL prompts. The targeted scans further localize these transitions in the intermediate state space and demonstrate that the identified directions are more informative than matched orthogonal directions. In particular, the handbag-count example provides an intuitive illustration of how neighboring intermediate states, under exactly the same remaining deterministic sampling process, can fall on opposite sides of a terminal decision boundary. These findings establish a concrete empirical connection between the boundary geometry characterized by our theory and meaningful changes in the outputs of real diffusion models.

4. Question: Does giving higher-risk trajectories more sampling steps improve generation quality?

A4. We sincerely thank the reviewer for this valuable suggestion. Motivated by this question, we conducted a preliminary proof-of-concept experiment to explore whether the sensitivity information studied in our theory could guide the allocation of sampling computation across trajectories.

Importantly, our goal here is not to introduce a new state-of-the-art sampler, but to examine whether trajectory-dependent sensitivity provides a useful signal for adaptive computation.

Building on the low-dimensional Heun study in Appendix C.4, we implemented a deliberately simple, vanilla allocation strategy based on the disagreement between two early clean-image estimates. After seven evaluations, trajectories with larger prediction drift received twelve additional evaluations, while the remaining trajectories returned their current estimates. We compared this strategy against random 7/19 allocation with exactly the same total computational budget, thereby isolating the potential benefit of trajectory-dependent allocation.

We evaluated the approach on native-resolution SDXL using 517 held-out GenEval prompts. Both GenEval and CLIP image–text cosine are reported as six-category averages. We also included Truncated Jump Sampling (TJS) as a fixed endpoint-decoding baseline and compared against standard sampling solvers.

Sampling ruleMean evaluationsGenEval ↑CLIP cosine ↑
Adaptive 7 or 1912.9944.16%0.2837
Random 7 or 1912.9942.10%0.2865
Fixed endpoint decoding (TJS)1347.34%0.2911
DDIM1349.61%0.2747
DPM-Solver++1350.86%0.2758
UniPC1351.30%0.2744
UniPC, lower budget1248.09%0.2727

Under the matched average evaluation budget, the simple drift-based allocation achieved a higher GenEval score than random allocation (44.16% vs. 42.10%), while CLIP cosine was slightly lower (0.2837 vs. 0.2865). We acknowledge that this preliminary strategy does not outperform established samplers such as UniPC, DPM-Solver++, or DDIM in GenEval. We do not intend to present it as a competitive replacement for existing sampling algorithms.

Nevertheless, we view this experiment as an encouraging initial indication that the trajectory-dependent sensitivity studied in our theory may contain useful information for allocating inference computation. It also reveals an important distinction: predicting whether a trajectory is sensitive to perturbations is not the same as predicting whether additional sampling computation will improve its final output. The current drift-based heuristic captures the former only indirectly and does not yet model the direction of improvement with respect to the requested generation attributes.

We believe this distinction points to an interesting opportunity for future research. Our theoretical framework provides a principled basis for characterizing where terminal sensitivity arises and how it propagates through the remaining diffusion dynamics. Building on this foundation, future algorithms could go beyond the present vanilla heuristic by jointly estimating terminal sensitivity, attribute-level correctness, and the expected benefit of additional computation.

We therefore view TRC primarily as a theory-grounded framework that opens up a new direction for sensitivity-aware sampling and selective computation, rather than as a claim that the current implementation already achieves state-of-the-art generation quality (this is our future work and may be the next article in this series).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.